: News
: Reviews
: Editorials

: Archives
: About
: Contact
: Advertising
: Privacy
: Links
: Forums




State of 3D: Q4 2007


8800 GT, 55nm, and Other Surprises


By Josh Walrath


            Today is the 8800 GT release day.  This is the first shot of a new graphics race that will heat up this holiday season, and for quite some time beyond that.  For now it is the 8800 GT’s chance to shine, as it is a new price performance champion with products already hitting shelves.  But the story for this upcoming month features far more than just the newly released 8800 GT, and we shall see some very interesting moves between both AMD and NVIDIA.  If you thought the past was cut-throat, I firmly believe that this November and beyond will see the marketing and product clashes of the past look to be pretty tame. 

The 8800 GT (aka G92)

            It has been a year since NVIDIA released the G80 based 8800 series of cards.  The GTX and GTS cards have been the top of the food chain when it comes to performance and features, and the further introduction of the lower priced GTS 320 helped cement its reputation as the video card of choice for enthusiasts willing to spend $250 and above.  AMD made some inroads with the R600 based HD 2900 XT, but design issues such as lackluster AA performance led to lower than expected sales.  When given the choice, most people bought the 8800 GTS over the 2900 XT.  But the G80 is a year old, and the chips being used for these parts are starting to look a bit long in the tooth when it comes to the cost of production as well as power consumption and heat.

            Rumors of the G92 chip began to start dropping before the weather started to get warm in Wyoming (which has an exceptionally long winter for being in the lower 48).  Confusion was the word of the season when speculating about the potential features of this chip, as well as what market it will be addressing.  It was not until about four weeks ago that we finally received concrete information regarding the chip.

            The G92 is essentially a shrink and optimization of the G80.  In macro-terms, the overall functionality is much the same as the G80 + NVIO1 chip.  But once we start digging, we see that there are some pretty significant differences… as well as a few unannounced surprises that will lead to interesting developments in the near future.  The G92 (as given by NVIDIA) is a 112 stream processor chip (officially 16 SPs less than a full G80).  Likely the 8th SP cluster is disabled, along with its texture unit.  Each texture cluster (made up of Texture Address and Texture Filtering units) is identical to that of the G84 chip, which powers the GeForce 8600 series of cards.  This means that it has doubled up the texture addressing units per cluster, but has left filtering untouched.

            The ROPS and memory support are perhaps one of the most interesting aspects of the design.  The G92 officially has 4 ROP clusters, each comprised of 4 ROPS.  This means overall there are 16 ROPS as compared to the 8800 GTS’s 20 ROPs and the GTX’s 24 ROPs.  The ROPs handle the anti-aliasing functionality as well, so right off the bat we see that while overall fillrate is still good, anti-aliasing performance will likely be slightly less when compared to its older siblings.  Each ROP cluster is attached to the memory crossbar, which is comprised of 4 x 64 bit memory busses.  The G92 officially supports a 256 bit memory bus, which is less than the 8800 GTS’s 320 bit bus, and the 8800 GTX which sports a nice 384 bit bus.

            The official specification calls for a card running at 600 MHz core, 1500 MHz shader clock, and 900 MHz memory.  It is also the first “official” PCI-E 2.0 compliant card, though there are rumors that the older RV630 and RV615 chips from AMD were PCI-E 2.0 compliant.  The card will have access to around 57 GB/sec of memory bandwidth, and connects to the visual world through 2 x dual link DVI ports.

            NVIDIA’s PureVideo 2 processor is also integrated into the G92 part, which allows it to further accelerate H.264 and VC-1 video.  It does not offload quite as much as the Avivo HD unit does, but it is pretty close.  NVIDIA has suffered some quality issues with their PureVideo HD, as there are some interesting artifacts when noise reduction is enabled in High Definition content.

            Going further into the G92, we see some interesting aspects that NVIDIA is not commenting on.  The chip itself is approximately 754 million transistors in size.  Considering the original G80 without the NVIO1 chip was approximately 681 million transistors.  At first glance one would assume that those extra transistors are about in line with all of the new features of the chip.  That might not be the case.  We can look back at NVIDIA’s previous jump from 130/110 nm to 90 nm to see what to expect.  The overall functional units of the 7800 GTX and 7900 GTX were identical, but the 7900 GTX was about 30 million transistors less than the older chip.  NVIDIA redesigned the chip so that it was able to take advantage of the new process node by removing a lot of redundant transistors, as well as “stage transistors” which would allow a chip to run faster due to fewer critical path limitations.  We can safely assume that NVIDIA did much the same with the jump from the 90 nm process down to the 65 nm process.  Once you consider that such an optimization could remove upwards of 80 to 100 million transistors, it becomes interesting to consider what all might be hiding under the hood.

            As mentioned above, NVIDIA probably has simply disabled portions of the G92 chip to get the features it has.  It is a pretty safe guess that the G92 does feature a full 128 SPs, 8 texture clusters, 6 ROP clusters, and the ability to address a 384 bit memory bus when all ROP clusters are enabled.  Obviously the full chip contains all of the same features as the full G80 chip, plus the additions of PV2, the NVIO1 chip, and some other interesting (yet unreleased) pieces.


Next:  More 8800 GT


If you have found this article interesting or a great help, please donate to this site.


Copyright 1999-2007 PenStar Systems, LLC.